Uncertainty Quantification In Surface Landmines and UXO Classification Using MC Dropout

TL;DR

Uncertainty quantification is critical for reliable landmine and UXO classification under challenging field conditions. The authors integrate Monte Carlo Dropout with a fine-tuned ResNet-50 to obtain predictive distributions by performing multiple forward passes during inference, treating the mean as the predicted class and the variance as a measure of epistemic uncertainty. They evaluate on a simulated four-class dataset under clean, adversarial (FGSM and PGD), and noisy conditions, showing that uncertainty rises with perturbations and can flag unreliable predictions. The work demonstrates the feasibility and potential utility of uncertainty-aware models in humanitarian demining and outlines directions for robust real-world deployment and extension to other detection/classification architectures.

Abstract

Detecting surface landmines and unexploded ordnances (UXOs) using deep learning has shown promise in humanitarian demining. However, deterministic neural networks can be vulnerable to noisy conditions and adversarial attacks, leading to missed detection or misclassification. This study introduces the idea of uncertainty quantification through Monte Carlo (MC) Dropout, integrated into a fine-tuned ResNet-50 architecture for surface landmine and UXO classification, which was tested on a simulated dataset. Integrating the MC Dropout approach helps quantify epistemic uncertainty, providing an additional metric for prediction reliability, which could be helpful to make more informed decisions in demining operations. Experimental results on clean, adversarially perturbed, and noisy test images demonstrate the model's ability to flag unreliable predictions under challenging conditions. This proof-of-concept study highlights the need for uncertainty quantification in demining, raises awareness about the vulnerability of existing neural networks in demining to adversarial threats, and emphasizes the importance of developing more robust and reliable models for practical applications.

Figures (5)

• Figure 1: Proposed modified ResNet-50 network architecture during training phase with last two layers unfrozen, modified final Fully Connected (FC) layer and a dropout layer before the final FC layer.
• Figure 2: Illustration of the testing phase, where multiple stochastic forward passes are performed on a trained model with the dropout layer enabled, enabling uncertainty estimation for the test sample.
• Figure 3: Predictive uncertainty on randomly selected clean test samples, with low uncertainty (i.e., high confidence) in (a) and (d), and higher uncertainty in (b) and (c).
• Figure 4: Model predictions and uncertainties with varying level of adversarial perturbations. (a) Clean image. (b) PGD attack ($\epsilon = 0.02$). (c) PGD attack ($\epsilon = 0.03$). (d) FGSM attack ($\epsilon = 0.001$) (e) FGSM attack ($\epsilon = 0.01$). (f) FGSM attack ($\epsilon = 0.05$).
• Figure 5: Predictive uncertainty on randomly selected noisy training samples, with low uncertainty in (a), (b), (c) and (d), and higher uncertainty in (e).